A clustering approach to injury severity in pedestrian-train crashes at highway-rail grade crossings

This research studied potential factors associated with pedestrian injury severity levels sustained in train-pedestrian crashes at highway-rail grade crossings (HRGCs) using the Federal Railroad Administration's ten-year data. The analysis focused on nonsuicide pedestrian crashes and took into consideration the unobserved heterogeneity. Latent class clustering (LCC) addressed unobserved heterogeneity by creating distinct subgroups with relatively homogeneous attributes within each subgroup. HRGC inventory variables were the basis for the LCC; the process split the dataset into five distinguished clusters. Binary logit models for each cluster and the complete data set were estimated. A generalized linear mixed model, based on the complete data set, allowed examination of the clustering and comparison of the modeling results. Findings provided justification for the use of LCC as the first step in accounting for unobserved heterogeneity. Different HRGC, pedestrian, and crash characteristics were associated with pedestrian injury severity across different clusters. Higher train speed was associated with more severe injury propensity, regardless of the conditions of the HRGCs. Other variables including freight train involvement, train hitting pedestrian, HRGCs with the absence of flashing lights, advance warnings, rural areas, lower visibility conditions, and older pedestrians increased pedestrian injury severity levels with varying effects in different clusters.